|
Abstract:
|
For most estimands recommended in the ICH E9 draft addendum "Estimands and Sensitivity Analysis in Clinical Trials", some amount of missing data can be encountered in the clinical trials practice. In this case, the assumptions about missing data must be consistent with the nature of the chosen estimand and appropriately incorporated in the estimation procedure. Pattern mixture models (PMMs) is a flexible analytical framework that have been proposed in the literature for this purpose. Several PMM variants use multiple imputation methodology to perform patient-level imputation under a variety of assumptions. They include such approaches as copy reference, jump to reference, unconditional reference, and imputation with delta adjustment. An alternative approach has also been proposed based on a longitudinal direct likelihood method with the imputation applied at the treatment group level rather than patient level. Its assumptions about missing data are similar to some multiple imputation variants mentioned above. We will discuss these methods and their underlying missing data assumptions as well as compare their performance.
|
